Proceedings, The Eleventh AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE-15)
Towards Generic Models of Player Experience
Noor Shaker1 Mohammad Shaker2 Mohamed Abou-Zleikha3
1
Center for Computer Games and Interaction Design at the IT University of Copenhagen, Copenhagen, Denmark
2
Joseph Fourier University, Grenoble, France
3
Audio Analysis Lab, AD:MT Aalborg University, Aalborg, Denmark
nosh@itu.dk,mohammadShakerGtr@gmail.com, moa@create.aau.dk
Abstract
adapt in other contexts. For instance, the probabilistic model
suggested in (Conati 2002; 2011) for affect detection in an
educational software, although accurate, can not be directly
used to predict affect in another educational system. Even if
we are to use the same methodology, we would need to go
through the whole process of data collection and model construction. This process is time and effort consuming and as
a result, the applicability of the systems built will be limited
to a specific case.
Each user has distinctive characteristics that could potentially be captured by a software. By investigating different systems and the way people interact with digital software, one could notice specific patterns and similarities of
a certain user behaviour across multiple systems. This suggests the potential of constructing generic estimators of UE
that could effectively predict user’s affect regardless of the
specific system the user is interacting with. One could ultimately envision a user-centric software that integrates different sources of information about the user (this could include
personality traits, physiological data, demographics, preferences, desires and behaviour) (Cooper, Reimann, and Cronin
2007) and is capable of accurately infering user’s affect and
consequently suggest personalised content. The concept of
context-independent users’ models is not new (Kobsa 2001;
Bartle 1996) yet very few examples can be found in the literature on providing a general representation of users (Alrifai et al. 2012; Heckmann et al. 2007) and this field has
not been, to the best of our knowledge, researched within
the games domain. This might be in part because of the difficulty in defining a shared model of UE that can be used
by different systems and because inferring users’ state from
data collected about the interaction with multiple systems is
not an obvious task.
Generic models of UE are particularly interesting in the
computer games domain where players engage in a rich
environment and where a shared model of user behaviour
is relatively easier to construct. Given the playing style in
one game, one could infer the player skill level in other
games that are not necessarily from the same genre and ultimately be able to use this information to personalise the
content (Bakkes, Tan, and Pisan 2012). It would therefore
be highly useful for a player model constructed from players’ data in one game to effectively predict the appeal of the
content of another game to a specific playing style.
Context personalisation is a flourishing area of research with
many applications. Context personalisation systems usually
employ a user model to predict the appeal of the context to
a particular user given a history of interactions. Most of the
models used are context-dependent and their applicability is
usually limited to the system and the data used for model
construction. Establishing models of user experience that are
highly scalable while maintaing the performance constitutes
an important research direction. In this paper, we propose
generic models of user experience in the computer games
domain. We employ two datasets collected from players interactions with two games from different genres where accurate models of players experience were previously built. We
take the approach one step further by investigating the modelling mechanism ability to generalise over the two datasets.
We further examine whether generic features of player behaviour can be defined and used to boost the modelling performance. The accuracies obtained in both experiments indicate a promise for the proposed approach and suggest that
game-independent player experience models can be built.
Introduction
The field of predicting user affect is rich with many interesting studies focusing on estimating user’s emotional states
while interacting with the system. The methodologies followed usually include collecting informative indicators of
user behaviour which are later employed by a machine learning technique to predict the appeal of a specific piece of
software to a particular user or group of users. Several behavioural features could be employed such as subjective, objective or predefined selected features gathered from the interaction with the system. The ultimate goal of most of these
systems is to construct accurate estimators of User Experience (UE) that could be later used to suggest content modifications or adjustments so that the system becomes aware
of the user progression and needs and ultimately able to provide personalised content.
Most of the studies reported however focus on analysing
and constructing models of UE in one setting. The models
are usually built with one specific case in mind and do not
usually scale; they can not be directly employed to predict or
c 2015, Association for the Advancement of Artificial
Copyright Intelligence (www.aaai.org). All rights reserved.
191
Modelling of User affects
This paper presents the first step towards this goal by investigating the feasibility of constructing accurate generic
estimators of players affect in the computer games domain. In particular, we investigate whether we can effectively capture players affect from generalised in-game behavioural features. We analyse two games from dissimilar
genres where accurate game-dependent estimators of player
Experience (PE) were previously constructed. We examine
the gameplay behaviour in both games and we investigate
the important features for predicting affect as selected by the
modelling mechanism. We further discuss whether the selected features can be generalised and applied in both games
and we consequently present experiments on constructing
generic models that effectively capture PE in both games.
The results suggest that accurate generic models of PE can
indeed be accomplished. To the best of our knowledge this
is the first attempt to construct generic models of UE from
behavioural features in the games domain.
Content personalisation is a crucial aspect in many applications where software are becoming more aware of their
users and more capable of adapting according to individual
needs (Viviani, Bennani, and Egyed-Zsigmond 2010). Rich
information about users and their interactions with systems
are collected to facilitate understanding users and their preferences for the purpose of providing “better” content. Such
systems can be employed in many different applications: serious games, e-commerce, e-learning, entertaining games,
etc.
User modelling (Kobsa 2001) plays an important role in
this type of environments and forms the basis for contextindependent personalisation (Niederée et al. 2004). Several studies in dissimilar domains suggest that providing
customised content improves user’s experience (Zakharov,
Mitrovic, and Johnston 2008; Conati 2002)
There is an abundance of studies presented in the literature on constructing computational models of emotion (Wöllmer et al. 2013; Calvo and D’Mello 2010; Ortony,
Clore, and Collins 1990; Conati 2002). There are also
game-specific theories about player emotion (Malone 1981;
Sweetser and Wyeth 2005; Koster 2004). Estimating affective and cognitive states in conditions of rich humancomputer interaction, such as in games, is a field of growing academic and commercial interest. Several studies with
varying success can be found on constructing models of
PE in different game genres (Yannakakis and Hallam 2009;
2006; Pedersen, Togelius, and Yannakakis 2010).
While the Player Experience Models (PEMs) constructed
in some of these studies achieved reasonable rates of accuracy (70-90%), they are still limited to predicting PE in
the specific game used to collect the data. There is only one
study we are aware of that investigate the construction of
generic models of PE from physiological data (Martı́nez,
Garbarino, and Yannakakis 2011). While this study suggests
a promise for the generalisation approach, its application
is limited to cases where physiological data is available or
can be easily acquired. Such data is usually not available in
the game domain where gameplay data is the most popular
source of information about player behaviour. Moreover, the
results obtained suggest that models with better accuracies
might be constructed if other modalities are considered.
Different studies on why people play games conducted
in an independent cross-genre research revealed a number
of key features of an optimal game experience that is generalised over game genres (Martı́nez, Garbarino, and Yannakakis 2011). These studies support our belief that player
behaviour in one game carry information about her general
behaviour in games and thereafter information about such
behaviour can generalise. If we are to prove this hypothesis,
there will be no longer a need to repeat the full process of
designing experiments, collecting data, and construing models of PE. The process could be optimised so that previously
construct models could be utilised for a new game and information coming from the user can be used to actively update
the model so that they become more accurate while the game
is being played.
Related Work
In what follows we review the relevant literature. In particular, we introduce previous work on emotion recognition
and on modelling UE with more emphasis on its application
within the computer games domain.
Emotion Recognition
In computer science, affective computing is the study and
development of methods that give the computers the ability to recognise and induce emotion and enable them to
interact with humans in human-like ways (Picard 1995).
The last three decades have witnessed increasing interest
in automatic human affect analysis. Several authors conducted extensive surveys of work in the machine analysis
of affective expressions (Sebe, Cohen, and Huang 2005;
Pantic and Rothkrantz 2003; Zeng et al. 2009).
In the games domain, affect induction is an essential part,
since most games can be tweaked in order to make the PE
more expressive and, thus, produce multimodal data that can
be analysed and classified (Scherer 2005). The literature defines three main modalities to capture affect (Yannakakis
and Togelius 2011). Objective measures such as data collected from e.g. body movement (Isbister, Schwekendiek,
and Frye 2011) and facial features (Hoque, McDuff, and
Picard 2012) among many others. Most of these measures
may be unsuitable for use in gaming context since they are
highly intrusive. Subjective measures on the other hand rely
on self-reports as the main indicator of affect. Despite their
simplicity in inferring user emotions, these methods have
been successfully implemented in a wide range of applications (Scherer 2005) including the work presented in this paper. Other studies relies on context-based features collected
from the interaction. These features are usually statistical
spatio-temporal features of interaction such as the amount
of time or the frequency of doing a certain activity. We use
such features in this work as the main method to capture behaviour
192
Method
Dataset 1: Super Mario Bros An executable version of
the software was uploaded online and participants were invited to play the game and answer the questionnaire. The
data was collected over a period of six months. Participants’
age covers a range between 16 and 64 years (31.5% females). The final dataset consists of a total of 273 unique
players who played 780 game pairs (1560 game levels). The
data was preprocessed to remove the pairs in which players reported unclear preferences (their answers were neither
or both equally), and after this step, the number of pairs remained were 597, 531 and 629 for engagement, frustration
and challenge, respectively. Several representative features
of player behaviour were extracted and Table 1 presents a
subset of these features. The full set contains 30 features
which can be found in (Shaker, Yannakakis, and Togelius
2012).
To test whether generic models could be constructed, two
dataset from games from dissimilar genres are used.
Super Mario Bros
Super Mario Bros is a very popular 2D platform game published by Nintendo. An open source clone of the game,
called Infinite Mario Bros (IMB) is used in our study. The
clone is modified to permit control over level generation and
thereafter provide different variations of content for players to experience and compare. The gameplay in Infinite
Mario Bros consists of moving the player-controlled character, Mario, through two-dimensional levels. Mario can walk
and run, duck, jump, and shoot fireballs. The main goal of
each level is to get to the end of the level. Auxiliary goals
include collecting as many coins as possible, and clearing
the level as fast as possible. The player is given three lives
to complete the game.
Dataset 2: First-Person Shooter A data collection event
was organised and advertised over social networks were people are invited to participate. The event was held at the university over two days and a total number of 62 students participated. Participants’ age covers the range between 18 and
24 years (12% females). Each player was asked to play at
least one pair (two game sessions, each last for two minutes) and she can play as many pairs as she wants. The
final dataset consists of 124 pairs and after removing the
pairs with unclear preferences 115, 111 and 112 pairs remained for engagement, frustration and challenge, respectively. Players’ behavioural features were extracted as indicators of players’ style. Table 2 presents a subset of these
features (the total set contains 23 unique features). The context features presented are the ones used to design the variations of the game content presented to the players (full details of the procedure can be found in (Shaker et al. 2013;
2015)).
First-Person Shooter: Sauerbraten
A second dataset from a First-Person Shooter (FPS) game is
used. The game is from a totally different game genre and
is built on the game engine Cube. The game in our experiments is played in single player mode and the goal is to
collect the highest score possible by traversing the arena for
two minutes, killing as many of the enemies as possible and
avoid being hit for a snapshot for one of the maps used). The
player is equipped with a weapon and she can collect other
types of weapons and resources. The weapons differ in their
accuracy, damage caused and shooting range. Every time the
player is killed, she looses one point and she is re-spawned
again as long as she still has time left to play.
Study Procedure
Player Experience Modelling
Two different datasets were collected containing different number of participants. Game surveys were conducted
to collect information about players’ interaction with the
games and their affective states. The same protocol is followed for data collection in both datasets. According to this
protocol, players are presented with a pair of two sessions
that differ along one or more aspects of game content. While
playing, detailed information about player behaviour and actions were recorded. After playing each pair, players were
asked to report their emotional/behavioural states following the four-alternative forced choice protocol that asks the
players to express their preference of the three states: engagement, frustration and challenge. Pairwise preferences
were adopted where the questionnaires presented are of the
form: “Which game was more E?” where E is the state
under investigation. The possible answers are: (1) game A
[B] was more/less E than game B [A] (2) both equally
or (3) neither. Full details about the procedure followed
can be found in (Shaker, Yannakakis, and Togelius 2012;
Shaker et al. 2013)
In the following sections, we describe the procedure followed to collect the data and the characteristics of each
dataset.
In order to construct powerful estimators of PE derived from
the in-game interaction we use Neuroevolution Preference
Learning (NPL). This approach has been used in the literature for modelling PE from preference data with better accuracies than other approaches such as support victor machine
and Bayesian learning (Yannakakis, Maragoudakis, and Hallam 2009). In the following we briefly describe the main
steps followed by this approach to build Player Experience
Models (PEMs), more details about the procedure can be
found in (Shaker, Yannakakis, and Togelius 2012).
Feature Selection
The first step when constructing models of PE is to analyse
the input space and select only the features that are relevant
for accurate prediction. For this purpose, we use Sequential Forward Selection (SFS). This is achieved by training
single-layer perceptrons (SLPs) and simple multi-layer perceptrons (MLPs) models through NPL.
Model Construction
SLPs and small MLPs can efficiently capture simple relationship between the features and reported affects. This re-
193
Table 1: Features extracted from Super Mario Bros data recorded.
Category
Time
Interaction
with items
Interaction
with enemies
Death
Miscellaneous
Context
features
Feature
tcomp
tplay
tjump
tlef t
tright
trun
tsmall
tbig
ncoin
ncoinBlock
kgoomba
kstomp
dtotal
dcause
nmode
njump
E
G
Gw
Ep
Description
Completion time
Playing duration of last life over total time spent on the level
Time spent jumping (%)
Time spent moving left (%)
Time spent moving right (%)
Time spent running (%)
Time spent in Small Mario mode (%)
Time spent in Big Mario mode (%)
Free coins collected (%)
Coin blocks pressed or coin rocks destroyed (%)
Times the player kills a goomba or a koopa (%)
Opponents died from stomping (%)
Total number of deaths
Cause of the last death
Number of times the player shifted the mode (Small, Big, Fire)
Number of times the jump button was pressed
Number of enemies
Number of gaps
Average width of gaps
Placement of enemies
Table 2: Extracted features from players data in the first-person shooter game.
Category
Time
Interaction
with items
Interaction
with enemies
Miscellaneous
Context
features
Feature
tlif e
tweapon
tshoot
tstill
tjump
texp
nhealth
narmour
ekill
phit
ehit
ndeath
sacc
E
Eskill
Wtype
H
R
Description
Duration of play
Time spent using weapons (%)
Time spent shooting (%)
Time spent not moving (%)
Time spent jumping (%)
Time spent using explosive weapons(%)
Health items collected (%)
Armours collected (%)
Number of times the player kills an enemy (%)
Number of times the player receives a hit from an enemy (%)
Number of times the player hits an enemy (%)
Number of times the player died
Shooting accuracy
Number of enemies
Skill level of enemies
Type of weapons including explosive and non-explosive weapons
Number of health items
Number of resources such as bullets and armors
lationship however is most likely to be more complex and
therefore after selecting the subset of relevant features, the
topology of the MLP models is optimised. This process
starts with small MLPs and the network topology is gradually complexified up to a certain predefined limit.
pendent runs using 3-fold cross validation. Each experiment
is repeated five times.
Results and Analysis
The above mentioned procedure was followed to construct
models of PE for the the two datasets. In the following, we
present the models constructed for each game and we discuss the generalisation procedure followed to obtain generic
models.
Experimental Setup
Parameter tuning tests were conducted to set up the parameters’ values for neuroevolutionary. As a result, we use
a population of 100 individuals and we run evolution for 20
generations. A probabilistic rank-based selection scheme is
used, with higher ranked individuals having higher probability of being chosen as parents. Finally, reproduction
is performed via uniform crossover, followed by Gaussian
mutation of 1% probability. In all of our experiments, the
data was first randomised to minimise the effect of playerspecific data on training and testing. The performance is calculated as the average classification accuracy in three inde-
Game-Specific Models
In previous studies, models of high accuracies were
constructed from subsets of selected features in both
datasets (Shaker, Yannakakis, and Togelius 2012; Shaker et
al. 2015). Table 3 presents the final set of features selected
following the PEM procedure discussed and the modelling
accuracies obtained. Notice that a baseline model in our case
would yield a prediction accuracy of 50%.
194
Table 3: Features selected from the set of extracted parameters for predicting engagement, frustration and challenge. The table
also presents the corresponding average performance (P̄avg ) and the maximum (Pmax ) values obtained. Context features appear
in bold.
Sauerbraten
Infinite Mario Bros
Engagement Frustration Challenge Engagement Frustration Challenge
Selected
phit
phit
tlif e
tcomp
tright
tplay
tstill
ehit
ndeath
ncoin
dtotal
njump
f eatures
Eskill
ekill
E
dcause
dcause
dtotal
E
tstill
Eskill
tsmall
kgoomba
ncoin
Wtype
Wtype
E
tplay
tright
texp
tweapon
tjump
Ḡw
Ḡw
narmour
ncoinBlock
G
Ep
tbig
njump
tlef t
trun
kstomp
njump
P̄avg
71.38%
80.91%
96.25%
67.18%
76.50%
74.03%
Pmax
77.19%
89.18%
99.09%
73.50%
83.00%
79.10%
Model Generalisation
Table 4: Average accuracies of the models (M ) when tested
on players’ data (D) from a different dataset. The models
built on the IMB dataset are tested on the FPS dataset and
vice versa.
Our first experiment aimed at investigating whether models
built on one dataset can be generalised to predict players’
affect on another dataset. To examine this, the accuracies of
the PEMs constructed from the IMB dataset are calculated
for the FPS data and vice versa. In order to allow such setup,
a transformation from one feature space (which is the input space to our PEMs presented in Table 3) to another is
needed. The features extracted fall in different ranges on the
game-dependency dimension. Some of the features can be
easily generalised such as the FPS features: the number of
enemies E, the number of enemies hit ehit and the average
amount of time spent in each life tlif e . These features carry
more or less the same information as the IMB features: the
total number of enemies E, the number of enemies killed
kgoompa and the time spent playing the game tcomp , respectively. Other features can be interpreted according to the
game such as the number of coins collected in IMB which, to
some extents, can be viewed as the number of health items
nhealth collected in the FPS game since they can be considered as different types of rewarding schemes. Some of
the features, however, such as the time spent using a specific
type of weapon in FPS, tweapon , the time spent in a big mode
in IMB, tbig , and the average width of gaps, Ḡw , are gamedependent and cannot be intuitively generalised. There are
a small number of these features and for testing purposes
we choose to fix their values to the average when evaluating
the models. After feature generalisation, The accuracies of
the models presented in Table 3 are recalculated on the data
from the other dataset and presented in Table 4.
The results indicate a wide verity in the performance.
The FPS models for predicting engagement for instance,
although relatively accurate on the FPS data, performed
poorly when evaluated on the data from players playing
SMB. On the other hand, the FPS models for predicting frustration performed very well on both datasets. It is interesting
to note that some of the models achieved better results when
evaluated on another dataset than the one used for model
construction. For instance, the IMB models for predicting
F P SM /SM BD
SM BM /F P SD
Engagement
31.47%
76.80%
Frustration
99.43%
55.61%
Challenge
72.01%
58.22%
engagement yield better results when evaluated on the FPS
data than those obtained from the IMB data. In general, it
appears that the differences in the performance obtained are
not contingent on predicted affective state nor on the prediction models. This suggests that specific models have better
generalisation ability than others. This is linked in part to the
set of the input features, their accurate generalisation and the
efficiency of the mapping process.
As discussed, although high accuracies are obtained is
some cases, poor performance is observed in others which
motivate conducting our second experiment.
Building Generic Models
In this experiment, we attempt to construct generic models from game-independent features as we believe that such
models will be more accurate, and ultimately more generic,
than those in the previous experiment. For this purpose, we
generalise the features in the two datasets and use the full
resultant set to construct new models. Practically speaking,
the features in each dataset are transformed into a more abstract level so that they represent boarder notion of players’
actions and interactions with the games’ objects. The definitions are chosen in a way that preserve the meaning of
the features while permitting their applicability to a wider
range of games. Feature generalisation is performed by going through all features from both games one-by-one and
trying to find a possible generalisation for it that can be
scaled to other games. This process resulted in a new set
of features that can be applied in both games. As not all fea-
195
ran another experiment where the feature space is generalised prior to model construction. PEM were then built with
the generic features as input and the two dataset combined
for training and testing. The experiment resulted in generic
models of high accuracies for the prediction of engagement,
frustration and challenge.
The experiments presented in this paper aimed at constructing models from players’ data in two different games.
Although the games are taken from dissimilar genres, they
share some characteristics that allowed feature generalisation. However, we might not be able to model PE from
data taken from players playing Tetris and SMB as there
will be very few features in common. The ultimate goal of
this line of research, is to construct models that could effectively work across many dissimilar game genres. This could
be achieved if richer information about the players is considered. Generic models can be constructed that incorporate
information about culture, knowledge, traits, demographics,
preferences, and many other objective and subjective modalities. Although more powerful machine learning techniques
might then become necessary, as the feature space will be
huge and since the mapping is not straightforward, the potential of such systems is undeniably significant.
Another potential application for the experiments presented in this paper is the possibility of providing a better
understanding of player behaviour in general that could infer and improve the design of user experiments. The generic
models constructed can be used as indicators of informative
features that impact affects and that designer should consider
when designing a new game or when examining PE.
The experiments presented in this paper are still preliminary indicators of the potential of the approach. The
method followed and the experimental setup used are motivated by previous research. However, since we are aiming at a different goal, it would be interesting to try other
modelling approaches for preference learning such as random forests (Abou-Zleikha and Shaker 2015), multivariate adaptive regression spline (Abou-Zleikha, Shaker, and
Christensen ) models or active learning approaches (Shaker,
Abou-Zleikha, and Shaker ) as well as experimenting with
different parameters tuned for the new data.
Also, the models constructed predict players’ affect across
three different affective states. With personalising PE as our
ultimate goal, one need to investigate whether previous approaches for personalising PE are still valid (Shaker, Togelius, and Yannakakis 2010; Bakkes et al. 2014) and analyse new approaches to select the best combinations of content features that optimises player’s experience across multiple dimensions.
Finally, we would like to point out that this paper draws
the outlines of this line of research, and thereof the word towards in the title, and it illustrate that there are indeed common gameplay behavioural patterns that can be captured and
that are scalable across games. The paper shed the light on
an important field of research in the game domain and motivates further investigations.
Table 5: Features selected from the set of generic features
for the prediction of engagement, frustration and challenge
for the combined dataset of IMB and FPS. The table also
presents the corresponding average performance (P̄avg ) and
the maximum (Pmax ) values obtained over five runs.
Combined Dataset (FPS + SMB)
Engagement Frustration Challenge
Selected
njump
tlif e
tlif e
Eskill
tstill
tstill
f eatures
E
ndeath
E
tstill
ekill
70.88%
79.81%
76.89%
P̄avg
Pmax
72.71%
80.99%
79.62%
tures can be scaled, some of the them are removed resulting
in a total of 14 generic features taken from both games. Such
features include for instance, the amount of time spent playing, moving or standing, the number of enemies killed, the
number of times the player died, etc.
The dataset used to construct the models is the result of
combining the data in the IMB and the FPS. The process of
feature selection and model construction is followed to buid
generic models and the experimental protocol presented in
Section is followed. The results obtained are presented in
Table 5.
Generic models with high prediction accuracies are obtained in all cases. The best models obtained are for predicting frustration with accuracy up to 81% while engagement is the hardest to predict (72.71%). The results are very
promising and suggest that generic feature of player behaviour can indeed be built and used to infer affect across
multiple games.
Notice that the preprocessing step of feature generalisation was necessary in the current study because we had the
features in both datasets already collected. However, the purpose of this experiment is to demonstrate that one could benefit from designing data collection experiments with generalised features in mind, as those can be used almost directly
to study behaviours in other similar games and the feature
generalisation stage will no longer be necessery.
Conclusions
In this paper, we presented and evaluated a method for building generic models of PE. For this purpose, we used two
datasets of player behaviour while interacting with games
from different genres. To test whether generic PEM can be
constructed, we employed previously constructed accurate
PE models and we examined their input space in an attempt to generalise them. These models were then tested
on another dataset than the one initially used to construct
the models. The results obtained from this experiment suggest a potential for the method and a possibility for better
prediction performance if the input space is optimised. As
our modelling method is heavily dependent on the input features selected, starting from generic features and allowing
the method to choose the relevant ones has the potential of
improving the performance. To further investigate this, we
Acknowledgments. This research was supported by the Danish Research Agency, Ministry of Science, Technology and Innovation project PLayGAle; project number: 274-09-0083.
196
References
Pedersen, C.; Togelius, J.; and Yannakakis, G. N. 2010. Modeling player experience for content creation. IEEE Transactions on
Computational Intelligence and AI in Games 2(1):54–67.
Picard, R. 1995. Affective computing. Perceptual Computing Section, Media Laboratory, Massachusetts Institute of Technology.
Scherer, K. 2005. What are emotions? and how can they be measured? Social science information 44(4):695–729.
Sebe, N.; Cohen, I.; and Huang, T. 2005. Multimodal emotion
recognition. Handbook of Pattern Recognition and Computer Vision 4:387–419.
Shaker, N.; Abou-Zleikha, M.; and Shaker, M. Active learning for
player modeling.
Shaker, N.; Shaker, M.; Abu-Abdallah, I.; Al-Zengi, M.; and
Sarhan, M. H. 2013. A quantitative approach for modelling and
personalizing player experience in first-person shooter games. In
UMAP Workshops.
Shaker, M.; Al-Zengi, M.; Abu-Abdallah, I.; Sarhan, M. H.; and
Shaker, N. 2015. Personalizing content generation in first person
shooter games through player modeling. In to appear.
Shaker, N.; Togelius, J.; and Yannakakis, G. N. 2010. Towards
automatic personalized content generation for platform games. In
Proceedings of the AAAI Conference on Artificial Intelligence and
Interactive Digital Entertainment (AIIDE). AAAI Press.
Shaker, N.; Yannakakis, G. N.; and Togelius, J. 2012. Crowdsourcing the aesthetics of platform games. IEEE Transactions on
Computational Intelligence and Games, Special Issue on Computational Aesthetics in Games.
Sweetser, P., and Wyeth, P. 2005. Gameflow: a model for evaluating player enjoyment in games. Computers in Entertainment (CIE)
3(3):3–3.
Viviani, M.; Bennani, N.; and Egyed-Zsigmond, E. 2010. A survey on user modeling in multi-application environments. In Advances in Human-Oriented and Personalized Mechanisms, Technologies and Services (CENTRIC), 2010 Third International Conference On, 111–116. IEEE.
Wöllmer, M.; Kaiser, M.; Eyben, F.; Schuller, B.; and Rigoll,
G. 2013. Lstm-modeling of continuous emotions in an audiovisual affect recognition framework. Image and Vision Computing
31(2):153–163.
Yannakakis, G., and Hallam, J. 2006. Towards Capturing and Enhancing Entertainment in Computer Games. In Advances in Artificial Intelligence, volume 3955 of Lecture Notes in Computer
Science. Springer Berlin / Heidelberg. 432–442.
Yannakakis, G. N., and Hallam, J. 2009. Real-time Game Adaptation for Optimizing Player Satisfaction. IEEE Transactions on
Computational Intelligence and AI in Games 1(2):121–133.
Yannakakis, G. N., and Togelius, J. 2011. Experience-Driven Procedural Content Generation. IEEE Transactions on Affective Computing.
Yannakakis, G. N.; Maragoudakis, M.; and Hallam, J. 2009. Preference learning for cognitive modeling: a case study on entertainment
preferences. IEEE Transactions on Systems, Man, and Cybernetics. Part A 39:1165–1175.
Zakharov, K.; Mitrovic, A.; and Johnston, L. 2008. Towards
emotionally-intelligent pedagogical agents. In Intelligent Tutoring
Systems, 19–28. Springer.
Zeng, Z.; Pantic, M.; Roisman, G.; and Huang, T. 2009. A survey of affect recognition methods: Audio, visual, and spontaneous
expressions. IEEE Transactions on Pattern Analysis and Machine
Intelligence 31(1):39–58.
Abou-Zleikha, M., and Shaker, N. 2015. Evolving random forest
for preference learning. In Applications of Evolutionary Computation. Springer. 318–330.
Abou-Zleikha, M.; Shaker, N.; and Christensen, M. G. Preference
learning with evolutionary-based multivariate adaptive regression
spline model.
Alrifai, M.; Gennari, R.; Tifrea, O.; and Vittorini, P. 2012. The
user and domain models of the terence adaptive learning system. In
International Workshop on Evidence-Based Technology Enhanced
Learning, 83–90. Springer.
Bakkes, S.; Whiteson, S.; Li, G.; Visniuc, G. V.; Charitos, E.; Heijne, N.; and Swellengrebel, A. 2014. Challenge balancing for personalised game spaces. In Games Media Entertainment (GEM),
2014 IEEE, 1–8. IEEE.
Bakkes, S.; Tan, C. T.; and Pisan, Y. 2012. Personalised gaming:
a motivation and overview of literature. In Proceedings of the 8th
Australasian Conference on Interactive Entertainment: Playing the
System, 4. ACM.
Bartle, R. 1996. Hearts, clubs, diamonds, spades: Players who suit
muds. Journal of MUD research 1(1):19.
Calvo, R. A., and D’Mello, S. 2010. Affect detection: An interdisciplinary review of models, methods, and their applications.
Affective Computing, IEEE Transactions on 1(1):18–37.
Conati, C. 2002. Probabilistic assessment of user’s emotions in educational games. Applied Artificial Intelligence 16(7-8):555–575.
Conati, C. 2011. Combining cognitive appraisal and sensors for
affect detection in a framework for modeling user affect. In New
Perspectives on Affect and Learning Technologies. Springer. 71–
84.
Cooper, A.; Reimann, R.; and Cronin, D. 2007. About face 3: the
essentials of interaction design. John Wiley & Sons.
Heckmann, D.; Schwarzkopf, E.; Mori, J.; Dengler, D.; and Kröner,
A. 2007. The user model and context ontology gumo revisited for
future web 2.0 extensions. Contexts and Ontologies: Representation and Reasoning 37–46.
Hoque, M.; McDuff, D.; and Picard, R. 2012. Exploring temporal
patterns in classifying frustrated and delighted smiles.
Isbister, K.; Schwekendiek, U.; and Frye, J. 2011. Wriggle: An exploration of emotional and social effects of movement. In Proceedings of the 2011 annual conference extended abstracts on Human
factors in computing systems, 1885–1890. ACM.
Kobsa, A. 2001. Generic user modeling systems. User modeling
and user-adapted interaction 11(1-2):49–63.
Koster, R. 2004. A theory of fun for game design. Paraglyph press.
Malone, T. 1981. What makes computer games fun? New York,
NY, USA: ACM.
Martı́nez, H.; Garbarino, M.; and Yannakakis, G. 2011. Generic
physiological features as predictors of player experience. Affective
Computing and Intelligent Interaction 267–276.
Niederée, C.; Stewart, A.; Mehta, B.; and Hemmje, M. 2004. A
multi-dimensional, unified user model for cross-system personalization. In Proceedings of the AVI 2004 Workshop on Environments
for Personalized Information Access, 34–54.
Ortony, A.; Clore, G.; and Collins, A. 1990. The cognitive structure
of emotions. Cambridge university press.
Pantic, M., and Rothkrantz, L. 2003. Toward an affect-sensitive
multimodal human-computer interaction. Proceedings of the IEEE
91(9):1370–1390.
197